Fundamentals Hit Bottom Will Rebound, AI is the New Focus (NVIDIA FY23Q4 Earnings Call)

On the early morning of February 23, Beijing time, NVDA.O released its Q4 FY23 financial report (as of January 2023) after the US stock market closed:

1. Key data vs market expectations

1. Revenue: NVIDIA US quarterly revenue of 6.051 billion US dollars, YoY-20.8%, in line with market expectations (6.02 billion US dollars).

2. Gross profit margin: Quarterly gross profit margin 63.3%, YoY-2.1pct, slightly lower than market expectations (64.3%).

3. FY24Q1 guidance:

(1) Revenue is expected to be USD 6.5 billion (±2%).

(2) GAAP and non-GAAP gross profit margins are expected to be 64.1% and 66.5% (+/-50 basis points), respectively.

(3) FY24Q1 capital expenditures are expected to be USD 350 million to USD 400 million, and FY24 capital expenditures are expected to be USD 1.1 billion to USD 1.3 billion.

For detailed financial report information, please refer to Dolphin Analyst's commentary "Surviving the Cycle Robbery, Meeting ChatGPT Again, and NVIDIA's Faith Rebirth".

2. Key content

1. Business fundamentals:

① Gaming business: Inventory adjustments are basically completed. Announced a 10-year partnership with Microsoft.US to bring Xbox PC game lineup to GeForce NOW, including blockbusters such as "Minecraft", "Halo", and "Flight Simulator".

② Data center: Some cloud service providers have suspended or adjusted construction plans and the impact on revenue in the China region. But the company expects growth to accelerate in the next period, thanks to the continued entry of H-100 into the market, demand for generative AI, and economic recovery.

2. AI: talk about AI throughout the report.

1) NVIDIA AI: It is the operating system of today's AI systems, which implements the entire process: data processing, learning and training, verification, and acceleration of inference.

2) NVIDIA DGX Cloud: Hosting it on the world's leading CSP will increase enterprise's capabilities in adopting NVIDIA AI Enterprise.

3) Hopper architecture: It has a Transformer engine, a new NVLink switch, and a new InfiniBand 400 gigabits/second data rate to achieve another leap in processing LLM. 4) Generative AI: Incredible ability and versatility. This is a very important moment for the computer industry. Every change in each platform, every turning point in the way computers are developed, is because it is easier to use, easier to program, and easier to obtain.

III. Conference Call Transcript

(I) Management Statement

Revenues for FY23Q4 were $6.05 billion, up 2% QoQ and down 21% YoY. Full-year revenues were $27 billion, flat YoY.

1. Data Center

(1) Revenue was $3.62 billion, down 6% QoQ and up 11% YoY. FY revenue was $15 billion, up 41%.

(2) Although some CSPs (Cloud Service Providers) paused or re-adjusted their construction plans at the end of the year, so the revenue from hyperscale customers did not meet the company's expectations. Nevertheless, it still achieved strong QoQ growth. Although expenditure adjustments usually reflect macroeconomic uncertainties as a whole, the company believes this is only a matter of time, as there is strong demand in the terminal market for GPU and AI infrastructure.

(3) The growth of network business was slightly lower than company expectations, due to weak demand for general infrastructure.

(4) The QoQ decline in total data center revenue was due to the impact of the epidemic on sales in China, which was basically in line with the company's expectations.

(5) As cloud computing applications continue to grow, the company is providing services to more and more CSPs, including Oracle and other GPU specialized enterprises. Last year, the revenue growth from CSP customers exceeded the overall revenue growth of the data center because more enterprise customers shifted to cloud-first. On the basis of the past four quarters, CSP customers contributed 40% of data center revenue.

(6) The new flagship product, H-100 GPU, is performing very strongly. In its second quarter alone, H-100 revenue was far higher than A-100 revenue (A-100 declined QoQ), demonstrating H-100's superior performance. The training speed of H-100 is 9 times that of A-100, and in inference based on Transformer-based LLM (large language models), its cluster speed is as high as 30 times. H-100's Transformer engine is launched just in time, and can serve the development and scaling of LLM's inference.

AI is at a turning point. OpenAI's ChatGPT has aroused the interest of the world, allowing people to experience AI first-hand, and demonstrating the possibility of generative AI. Whether it is generating marketing copy, summarizing documents, creating images for advertisements or video games, or answering customer questions, these new neural network models can improve productivity in large-scale tasks. The generative AI application will help every industry to complete more work faster. The generative LLM with over 100 billion parameters is the most advanced neural network in the world today, and NVIDIA's expertise spans AI supercomputers, algorithms, data processing, and training methods, enabling it to provide these capabilities to enterprises. The company is looking forward to helping clients achieve generative AI.

(7) In addition to collaborating with each major CSP, the company is also working with many consumer internet companies and startups, driving strong growth in data centers, which will accelerate later this year.

(8) This quarter, the company made significant progress in one of its largest vertical industries—the financial services sector. First, in collaboration with Deutsche Bank, it accelerated the application of AI and machine learning in the financial services industry. The company and Deutsche Bank jointly developed a series of applications, including virtual customer service personnel, voice AI, fraud detection, and bank process automation, using NVIDIA's entire computing stack (including NVIDIA AI Enterprise) internally and in the cloud. Second, it achieved leading AI inference results in applications such as asset price discovery, a critical benchmark for the financial services industry.

(9) In terms of networking, driven by AI, demand for the latest generation of InfiniBand and HPC-optimized Ethernet platforms continues to grow. The scale of the basic model of generative AI continues to grow exponentially, driving demand for high-performance networks to expand multi-node accelerated workloads. InfiniBand provides unparalleled performance, latency, and network intra-compute capabilities, making it the best choice for efficient cloud scale and generative AI.

For small-scale deployments, NVIDIA will bring its full stack acceleration stack expertise and integrate it with the world's most advanced high-performance Ethernet architecture. This quarter, driven by demand from cloud computing, enterprise, and supercomputing customers, the Quantum 2 40 Gbps platform launched smoothly, and InfiniBand led the growth. In terms of Ethernet, as customers transition to higher speeds, next-generation adopters, and switches, the development of the Spectrum 4 network platform at 40 Gbps per second is strong.

(10) The company is still focused on expanding its software and services. NVIDIA AI Enterprise 3.0 was released, supporting more than 50 NVIDIA AI frameworks and pre-training models, as well as new workflows for contact centers, intelligent virtual assistance, audio transcription, and network security. Upcoming products include the NeMo and BioNeMo LLM services, which are currently being communicated with customers in the initial stage.

(11) Accumulation of technological breakthroughs has pushed AI into a turning point. The versatility and capabilities of generative AI have prompted businesses around the world to urgently develop and deploy AI strategies. However, the infrastructure for AI supercomputers, model algorithms, data processing, and training techniques for AI remains an insurmountable barrier for most people. Today, the company announces the next level of its business model to help bring AI within reach of every enterprise customer.

The company is collaborating with major CSPs to offer NVIDIA AI cloud services. These services are provided directly by NVIDIA through a market partner network, hosted in the world's largest cloud. NVIDIA AI services provide easy access to the world's most advanced AI platform for enterprises, as well as access to the storage, network, security, and cloud services provided by the world's most advanced cloud. Customers can participate in NVIDIA AI cloud services on AI supercomputers, accelerated library software, or pre-trained AI models. NVIDIA DGX is an AI supercomputer and the blueprint for AI factories being built around the world. Building an AI supercomputer is difficult and time-consuming. Today's announcement of NVIDIA DGX Cloud is the fastest and easiest way to have your own DGX AI supercomputer. All you need to do is open your browser. NVIDIA DGX Cloud is provided through Oracle Cloud infrastructure, Microsoft Azure, Google GCP, and other platforms that will be launched soon.

In the AI platform software layer, customers can access NVIDIA AI Enterprise for training and deploying LLM or other AI workloads. At the pre-trained generative AI model layer, two customizable AI models, NeMo and BioNeMo, will be provided to enterprise customers who wish to establish proprietary generative AI models and services for their business. Through this new business model, customers can participate in NVIDIA's comprehensive AI computing on their private clouds or any public cloud. More details about NVIDIA's AI cloud services will be shared at the upcoming GTC, so stay tuned.

2.Gaming

(1) Gaming revenue was US$1.83 billion, up 16% quarter over quarter and down 46% year over year. For the fiscal year, revenue was US$9.07 billion, a 27% decrease. The QoQ growth was mainly due to strong demand for the 40 series GeForce RTX GPUs. The YoY decline reflects the impact of channel inventory corrections, but most of the impact has ended. Demand in most regions was solid during the seasonally strong fourth quarter. Although China was somewhat affected by pandemic-related disruptions, the company is encouraged by early signs of recovery in that market.

(2) Players responded enthusiastically to the new RTX 4090, 4080, and 4070 Ti desktop GPUs, and inventory at many retail and online stores sold out quickly. The flagship product, RTX 4090, quickly climbed to the top of the AI architecture popularity chart on Steam, reflecting players' desire for high-performance graphics.

Earlier this month, the first phase of Ada-based gaming laptops was launched for retail, achieving NVIDIA's largest generational leap in performance and power consumption to date. It brings high-end GPU performance to 14-inch laptops for the first time, which is a rapidly growing segment of the market that was previously limited to basic tasks and applications.

In addition, thanks to the power efficiency of fifth-generation Max-Q technology, the 90-level GPU (the best-performing model) was applied to laptops for the first time.

In summary, the RTX 40 series GPUs will power more than 170 gaming and creator laptops, preparing for back-to-school season. Now, there are over 400 games and applications that support NVIDIA RTX technology for real-time ray tracing and AI-driven graphics. The hallmark of the AI architecture is DLSS 3, or the third generation of AI-driven graphics, which has greatly improved performance. The cutting-edge game "Cyberpunk 2077" recently joined DLSS 3, boosting frame rate performance at 4K resolution three to four times.

(3) The GeForce NOW cloud gaming service continues to expand in terms of users, performance, and more. It now has over 25 million members in more than 100 countries. Last month, it enabled RTX 4080 graphic horsepower in a new high-performance ultimate member level. Ultimate members can stream from the cloud at 240 frames per second and have full ray tracing and DLSS 3.

Yesterday, it was announced that a 10-year partnership with Microsoft has been established to bring Xbox PC games to GeForce NOW, including hit titles such as "Minecraft," "Halo," and "Flight Simulator." After Microsoft's acquisition of Activision Blizzard, games such as "Call of Duty" and "Overwatch" will also be added.

(3) Professional design visualization solutions:

(1) Revenue was USD 226 million, a 13% increase from the previous quarter and a 65% decrease from the same period last year. The vertical advantage of desktop workstations in the automotive and manufacturing industries drove quarterly growth. The YoY decrease reflects the impact of channel inventory corrections, which are expected to end in the first half of this year.

(2) Interest in Omniverse continues to grow, with nearly 300,000 downloads and 185 connections to third-party design applications to date. The latest release of Omniverse features many new functionalities and improvements, including support for 4K, real-time path tracing, AI-driven Omniverse search through large untagged 3D databases, and Omniverse cloud containers for AWS.

(4) Autonomous driving technology and solutions:

Revenue reached a record USD 294 million, a 17% increase from the previous quarter and a 135% increase from the same period last year. The QoQ increase was mainly driven by AI automotive solutions, and new projects from electric vehicle and traditional OEM customers also contributed to growth. Annual revenue was USD 903 million, a 60% increase.

At CES, it was announced that a strategic partnership with Foxconn has been established to jointly develop autonomous driving and self-driving car platforms. This will provide mass-produced scale to meet the growing demand for NVIDIA Drive platforms. Foxconn will use NVIDIA's Drive, Hyperion computing, and sensor architecture in its electric vehicles, making it a tier one manufacturer producing electronic control units based on NVIDIA's Drive Orin for global customers. This quarter marks an important milestone: the NVIDIA Drive operating system has obtained TÜV SÜD safety certification. TÜV SÜD is one of the most experienced and rigorous assessment agencies in the automotive industry. With leading industry performance and functional safety, the NVIDIA platform has achieved higher standards required for autonomous transportation.

（二）FY24Q1 Guidance

The strong growth of the data center and gaming businesses will drive growth for the four major market platforms and achieve sequential growth.

Revenue is expected to be $6.5 billion (+/-2%).
GAAP and non-GAAP gross margins are expected to be 64.1% and 66.5% (+/-50 basis points), respectively.
GAAP operating expenses are expected to be $2.53 billion, and non-GAAP operating expenses are expected to be $1.78 billion.
GAAP and non-GAAP other income and expenses are expected to be $50 million (excluding gains and losses from non-affiliated spin-offs).
GAAP and non-GAAP tax rates are expected to be 13% (+/-1%, excluding any discrete items).
FY24Q1 capital expenditures are expected to be $350 million to $400 million, and FY24 capital expenditures are expected to be $1.1 billion to $1.3 billion.

（三）Q&A

Q: Monetization contribution of software and cloud strategy to company's business model in the future quarters

A: In FY23Q4, the company made good progress in collaborating with partners, increasing the number of partners, and promoting the development of software business. The revenue from the software business is in the billions, reaching a record level in Q4.

Essentially, NVIDIA AI is the operating system of today's AI system, which accelerates the entire process: data processing, learning training, validation, and inference. It can run in every cloud, can be run locally, supports every framework and model, and can be accelerated anywhere. Using NVIDA AI-accelerated software can improve the efficiency and cost-effectiveness of the entire machine learning process, thereby saving money.

Today, the company announced the launch of NVIDIA DGX Cloud and will host it on the world's leading CSP, which will enhance the enterprise's ability to adopt NVIDIA AI Enterprise and expand people's adoption of machine learning. Although it is not widely deployed in enterprises, the company believes that hosting everything in the cloud: infrastructure, operating system software, and pre-trained models can accelerate the application of generative AI in enterprises.

The company is excited about the new expansion of the business model: NVIDIA DGX Cloud, which will accelerate the adoption of software.

Q: FY24Q1, Guidance on Data Centers

A: Expect strong sequential and year-on-year growth in data centers, with FY24Q1 expected to accelerate its year-on-year growth. Q: The computing power of generative AI is very high, will this limit the market size to a few super-large companies? If the market is very large, will it attract Cloud A6 or other accelerators to join the competition?

A: (1) Is the market size limited to a few super-large companies?

First of all, LLM is called a large-scale language model because it is quite large. However, in the past 10 years, the company has accelerated AI processing by 1 million times. Moore's Law, at its best, can achieve 100 times in 10 years. By introducing new processors, systems, interconnections, frameworks and algorithms, and working with data scientists and AI researchers to develop new models, the company has increased the processing speed of LLM by one million times over the entire 10-year span.

It used to take several months, but now it takes about 10 days. Of course, you still need a large infrastructure. For this, the company is also introducing the Hopper architecture. Hopper has a Transformer engine, a new NVLink switch, and a new data rate of 400 gigabits per second for InfiniBand to achieve another leap in processing LLM.

Therefore, by combining NVIDIA DGX supercomputers with NVIDIA DGX Cloud, it is easier for people to access and use this infrastructure, and truly make this technology and capability quite easy to obtain by accelerating training capabilities.

Secondly, a considerable number of LLM or basic models must be developed. Different countries have different cultures, and their knowledge systems are different. Whether it is radiology, biology, or physics, each field needs its own basic model. Thanks to LLM, there is now a precedent that can be used to accelerate the development of all these other fields.

Third, many companies have their proprietary data. The most valuable data in the world is proprietary and belongs only to the company, and these proprietary data will be used for the first time to train new AI models. Therefore, the company's strategy and goal is to put DGX infrastructure in the cloud to provide this capability to every company that wants to create proprietary data and proprietary models.

(2) Competition

The company has been in competition for a long time, but its methods and computing architecture are quite different:

First, it is generic. This means it can be used for training, inference, and adapting to different types of models; it supports every framework, every cloud; it is ubiquitous, from the cloud to the private cloud, from the cloud to the enterprise, all the way to the edge; it can be an autonomous system, allowing developers to develop their own AI models and deploy them anywhere.

Second, AI itself is not an application. In fact, it has a pre-processing part and a post-processing part to turn it into a program or service. Most people don't talk about pre-processing and post-processing, maybe because it's not that interesting. However, it has been proven that pre-processing and post-processing often consume half or two-thirds of the entire workload. The company can accelerate the entire end-to-end pipeline: preprocessing (data ingestion-data processing) - post-processing, not just speeding up half of the pipeline. If you only accelerate half of the workload, even if you pass it immediately, the speed will only be doubled. But if the entire workload is accelerated, it can be accelerated by 10 times, 20 times, or 50 times. When you hear about NVIDIA accelerating applications, you will often hear about 10x, 20x, and 50x acceleration, because the company not only accelerates the deep learning part, but also uses CUDA to accelerate the entire process end-to-end.

The universality of the company's acceleration platform, including its existence in every cloud, from the cloud to the edge, makes it more accessible and easier to use, and more unique. Most importantly, for all CSPs, due to the high utilization efficiency of the acceleration platform - which can accelerate end-to-end workloads and achieve good throughput - the company's architecture has achieved the lowest operating costs, which cannot be matched by competitors.

Q: Jensen mentioned that ChatGPT is a turning point similar to the iPhone, so what changes have occurred in the company's communication with ultra-large-scale and large-scale enterprises after ChatGPT? Secondly, what has happened in the past few months regarding the growth prospects of Hopper with its transformative engine and Grace with its high-bandwidth memory?

A: (1) ChatGPT

ChatGPT is a wonderful work. The OpenAI team created a model that can be nested internally. The versatility and ability of ChatGPT surprised everyone. Specifically, people were surprised that a single AI model could perform tasks and skills that had never been trained before. This language model can input human language and output Python, output Cobalt (a language that few people even remember), and output Python for Blender (a 3D program). So, ChatGPT is a program for writing programs for another program. The world now realizes that perhaps human language is a perfect computer programming language, and NVIDIA has democratized computer programming for everyone (that is, it is easier to access and deploy computer programming), and almost anyone can use human language to explain specific tasks to be performed. This new computing platform can accept any prompts and requests from humans and translate them into a series of instructions. It can execute these instructions directly or wait for the user to decide whether to execute them.

Therefore, this type of computer application is completely revolutionary. It realizes the democratization of programming and enables every CSP or, frankly, every software company to use it. Because this is an AI model that can write programs for any program, every software programmer who is developing software is either shocked or actively researching models similar to ChatGPT to integrate them into their own programs or services.

(2) Hooper and Grace

In the past 60 days, significant breakthroughs have been made in building the AI infrastructure for Hopper, as well as using Hopper and Ampere to infer LLM work. Therefore, there is no doubt that regardless of the company's view of this year when it started, there have been significant changes in this view in the past 60 or 90 days. Q: How will the business mix of data centers evolve into systems and software over the next 2-3 years, and how will this affect profit margins?

A: Data center business is only conceptually GPU business, because the company actually sells a fairly large computing panel consisting of 8 Hoppers or 8 Amperes to CSP, which is connected to NVLink switch. This panel basically represents 1 GPU. It consists of 8 chips connected into 1 GPU with very high-speed interconnects between chips. In fact, the company has been working on multi-chip computers for quite some time. Therefore, GPU is actually HGX GPU, which is 8 GPUs, and the company will continue to maintain this approach. What truly excites CSP is hosting the company's infrastructure and providing services. Currently, the company cooperates directly with a large number of enterprises, including 10,000 AI startups worldwide, and enterprises from every industry. All partnerships are very much looking forward to deployment in the cloud and internal, or at least deployment in the multi-cloud.

Therefore, by enabling NVIDIA DGX and NVIDIA infrastructure to achieve full stack in their cloud, the company is attracting CSP customers. For the company, NVIDIA will be the best cloud computing AI sales force in the world. For customers, they have access to the most advanced real-time infrastructure, a team with infrastructure-accelerated software-NVIDA AI Open Operation System-AI model. In a single entity, customers can obtain expertise across the entire span. This is a great model for both CSP and NVIDIA.

The company will continue to promote DGX AI supercomputers, but it does take time to establish AI supercomputers internally. Now the company has already obtained a lot of things in advance, so that customers can start and run as quickly as possible.

Q: Most of the focus is currently on text, but clearly some companies have done a lot of training on video and music. It may take at least 10,000 GPUs in the cloud to train these large models, and several tens of thousands of GPUs may still be needed to infer a widely deployed model. Therefore, incremental TAM (Total Addressable Market) can easily reach hundreds of thousands of GPUs, that is, it is easy to reach a scale of tens of billions of dollars. What impact does this have on the TAM (potential market size of $300 billion in hardware and $300 billion in software) given last year, and how do you view the new TAM?

A: These figures are still very good anchors. The difference is that due to the incredible capabilities and versatility of generative AI, as well as all the fusion breakthroughs that occurred in the middle and end of last year, the company may reach this TAM earlier. There is no doubt that this is a very important moment for the computer industry. Every change of platform, every inflection point in the development of computer methods, is because it is easier to use, easier to program, easier to get. This happened in the personal computer revolution, this happened in the Internet revolution, and this happened in the mobile cloud. Thanks to iPhone and App Store, there are 5 million mobile cloud applications available and the number is still increasing. At that time, there were no 5 million mainframe programs, workstation programs, or PC programs. Now, it is so easy to develop and deploy programs, partly in the cloud and partly on mobile devices, and easily distribute through application stores. The same thing is happening in AI now.

ChatGPT has reached 150 million people unprecedentedly in 60-90 days, and people are using it to create various things, which is quite unusual. At present, what everyone sees is just a tide of new companies and new programs appearing. Without a doubt, this is a new era of computing. As a result, TAM will be easier and faster to achieve than before.

Q: Previously mentioned that "H-100 revenue is higher than A-100", is this the overall situation or based on a specific point in time (such as after the shipment of two quarters)?

A: As early as the initial shipment of H-100 in FY23Q3, Q4H-100 shipments surged, which means that many CSPs want to launch and operate H-100 in Q4. Therefore, the number of Q4A-100s is less than that of H-100s. The company will continue to sell A-100 and H-100, but for Q4, H-100 sales is very strong.

Q: Mercedes-Benz claims that MB Drive software revenue may be in the low billions of euros by the mid-decade and in the tens of billions of euros by the end of the decade, while NVIDIA splits software revenue in half. Is this software revenue generated by the Mercedes-Benz transaction and is it on a similar timeframe as Mercedes-Benz?

A: Mercedes-Benz has communicated with the company about software collaboration, which is divided into two parts: what can be done with Drive and what can be done with Connect, which will be approximately 10 years. This is consistent with the company's idea of sharing revenue as a long-term partner over time.

Mercedes-Benz is the only large luxury brand that installs a rich sensor array and AI supercomputer on every car from entry-level to highest-end, which establishes an upgradable installation base for every car and a continuing update in the future. If the entire Mercedes-Benz fleet is fully programmable, it can be OTA, thereby creating revenue.

Q: Is there research on new workloads or programs to drive the next level of demand?

A: The company will share its new programs and workloads with everyone at GTC. Stay tuned!

There are three reasons for constantly introducing new programs:

First, NVIDIA is a multi-domain acceleration computing platform, not completely universal like CPU (98% control function + 2% math -> completely flexible). NVIDIA is an acceleration computing platform that works with CPUs to unload heavy and highly paralyzed computing units. However, NVIDIA covers particle systems, fluids, neurons, and computer graphics, becoming a multi-domain platform. Second, the installation base is very large. It is the only accelerated computing platform that is compatible with every cloud architecture, including PCs, workstations, game consoles, cars, and enterprise internals. The versatility, convenience, acceleration performance, and ecosystem of the NVIDIA platform are loved by developers who have developed special programs. More importantly, they like the fact that they can reach a large number of users after developing software, which is also the reason why the company continues to attract new programs.

Third, the speed of CPU computing progress has greatly slowed down. As early as the first 30 years of my career (Jensen), the performance increased by 10 times every 5 years with approximately the same power consumption to promote. Now, the speed of this continuous progress has slowed down. But demand is still strong, and due to sustainability, it cannot withstand the continuous increase in power consumption. Acceleration can reduce the amount of electricity used in any workload.

These reasons have prompted people to use accelerated computing to promote companies to continuously create new programs.

Q: The construction plan of Data Center in January seemed to be a bit weak, but the FY24Q1 and full-year guidance of Data Center both show YoY accelerated growth. What is driving the accelerated growth (whether it is based on H-100, generative AI sales, or new AI service models), and how do you see the situation in the enterprise vertical field?

A: The Data Center will achieve MoM growth and YoY accelerated growth in FY24Q1.

The driving factors are as follows:

First, multiple products are entering the market. Now, H-100 has entered the market, and it is expected to launch new products in the second half of this year, which are sometimes powered by GPU computing and networking.

Second, customers have a strong interest in generative AI, including CSPs, enterprises (including start-ups), and are expected to drive revenue growth.

Finally, with the improvement of the economy, AI and accelerated computing are very important to enterprises, and the cloud can promote the development of enterprises.

Q: Hopper is 6 times the performance of Transformer. How to expand this scale and how much of it reflects that it will just be a larger hardware expenditure?

A: As AI (i.e. intelligent production) will become a manufacturing industry, the number of AI infrastructures will increase worldwide. In the past, people only made physical commodities; in the future, almost every company will make intelligent forms of software products. The data center only does one thing: adjust the data and create new models. After the raw materials enter, the infrastructure processes them, produces or improves products, and this is the "factory." Part of the AI factory will be hosted in the cloud, part inside the enterprise; some are large, some are super large, and some are small.

Q: The ability to train models has increased by a million times in the past 10 years. In what time frame will customers with these LLMs need 100 times the complexity in the coming years?

A: In the next 10 years, it is hoped that through new chips, new interconnections, new systems, new operating systems, new distributed computing algorithms and new AI algorithms, as well as working with developers to launch new models, AI can be accelerated by another million times. This is also one of the reasons why NVIDIA is not just a chip company. The problems the company aims to solve are just too complex, and the whole stack must be considered (starting from the chip, going through software, and all the way to the data center). NVIDIA can think and innovate across the entire stack. Therefore, this is indeed a playground for computer scientists.

In the next 10 years, people will see truly huge breakthroughs in AI models and platforms. At the same time, as a result of these breakthroughs and the adoption of AI, people will see AI factories everywhere.

Dolphin NVIDIA Historical Articles:

In-depth

June 6, 2022, "Did the US Stock Market Shock Mistaken Apple, Tesla, and NVIDIA?"

February 28, 2022, "NVIDIA: High Growth is Not False, but the Price-Performance Ratio Still Lags Behind"

December 6, 2021, "NVIDIA: Valuation Cannot Be Propped Up by Imagination Alone"

September 16, 2021, "NVIDIA (Part 1): How Did the Chip Genius Increase Twentyfold in Five Years?"

September 28, 2021, "NVIDIA (Part 2): The Double Drive is Gone, Will Davis Make a Double Kill?"

Earnings Season

February 23, 2023, Earnings Review, "Surviving the Cycle Disaster and Meeting ChatGPT, NVIDIA Faith Returns"

November 18, 2022, Conference Call, "Can the High Inventory that Continues to Rise be Digested Next Quarter? (NVIDIA FY2023Q3 Conference Call)"

November 18, 2022, Earnings Review, "NVIDIA: Profit Fell Two Thirds, When Will the Turning Point Come?"

August 25, 2022, Conference Call, "How Does Management Explain the "Flash Crash" Gross Margin? (NVIDIA FY2023Q2 Conference Call)" Earnings Report Reviews: